feat: per-machine hardware profiles — detect the card, fit the box (13-task plan)#110
Merged
Conversation
…nment + markdownlint + rubric gate) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
#81, t7) Add lobes assess --probes: a role that is /health-healthy but semantically wrong now FAILS its probe, instead of only /health being checked. - cortex: known-answer question, temperature 0, exact-match on a forced short completion (POST /v1/chat/completions). - embedder: embeds a sentence + paraphrase + unrelated string; PASS iff cos(sentence, paraphrase) > cos(sentence, unrelated). Enforces a hard per-request timeout and treats a timeout as FAIL, not skip — this is the shape that would have caught the sm_110 FLASH_ATTN hang (request accepted, never answered). - reranker: PASS iff the one relevant document ranks first by relevance_score (POST /v1/rerank). Each probe is a pure function of (url, model) over the existing stdlib _post()/_get() transport, monkeypatchable exactly like the existing run_correctness() probes — no new dependency. The CLI wiring resolves each role's own endpoint via the existing role registry (lobes.roles.role_registry_from_env), the same builder lobes capabilities/ lobes measure use, so --probes stays read-only (no compose/env writes). 29 new tests in tests/test_role_probes.py cover pass/fail/timeout/malformed paths for all three probes (mocked transport, no network/GPU), the run_role_probes/render_role_probes glue, and the CLI surface (--json, --role filter, --timeout, no --apply, never touches docker). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
One module per card under lobes/machines/ — each a CardStrategy owning its own detection signature, per-role knobs and per-knob provenance — plus a small stdlib registry (register/unregister/get/names/strategies/detect) and a shared _traits module so sm_110 pooling quirks compose without copy-paste. - thor is now load-tested: cortex kv_cache_dtype=auto, embedder/reranker on TRITON_ATTN (+ reranker enforce_eager) via the shared SM_110 trait; the legacy single-model row no longer claims the contradicted flashinfer. - detect() is honest (returns None on no match); generic never auto-matches. - registration order == detection precedence (spark before blackwell); pinned off isort so it is never alphabetised. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
MachineProfile / MACHINE_PROFILES / machine_profile / detect_machine / machines_as_dicts stop carrying a per-chip table and are rebuilt from lobes.machines on each access. MACHINE_PROFILES is served via module __getattr__ (PEP 562) so a chip registered after import is reflected with no edit here. All legacy behaviour (and its callers in switch/benchmark) is preserved; detect_machine keeps its silent generic fallback over the registry's honest None resolver. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
Covers the three acceptance criteria: a synthetic chip registered at runtime flows through detection, profile resolution, MACHINE_PROFILES and knob rendering with zero edits elsewhere; the sm_110 trait is shared not copied (board override still wins); thor carries the measured knobs and its legacy row drops flashinfer while detect_machine still falls back to generic. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…les derived Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
Every machine-dependent knob the Thor (sm_110) box hand-edits now flows
through .env with today's exact rendered behavior as the default (GB10
sees zero behavior change):
- vllm-primary --kv-cache-dtype -> ${PRIMARY_KV_CACHE_DTYPE:-fp8}
- vllm-embed / vllm-rerank gain --attention-config, defaulting to
"auto" (verified against vllm/config/attention.py: the string
"auto" is vLLM's own automatic-selection sentinel, so this is
byte-equivalent in effect to today's flagless render, not a guess)
- vllm-rerank gains an enforce-eager toggle, defaulting to the
explicit --no-enforce-eager no-op (vLLM's enforce_eager field uses
argparse.BooleanOptionalAction, so both --enforce-eager and
--no-enforce-eager are always-valid tokens, never a conditionally-
omitted flag)
- the legacy single-model template's --kv-cache-dtype gets the same
treatment (VLLM_KV_CACHE_DTYPE)
Verified with `docker compose config`: the default render is unchanged,
and setting only PRIMARY_KV_CACHE_DTYPE=auto, EMBED_ATTENTION_BACKEND=
TRITON_ATTN, RERANK_ATTENTION_BACKEND=TRITON_ATTN, RERANK_ENFORCE_EAGER=
--enforce-eager reproduces /home/thor/.lobes/docker-compose.yml's
vllm-primary/vllm-embed/vllm-rerank commands byte-for-byte.
Deviation (binding, per the task brief): MULTIMODAL_ATTENTION_BACKEND on
the senses/vllm-multimodal gear is left exactly as-is, with a comment
pointing at #109 — its removal is gated on a GB10
live-verification, not this task.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…force-eager); MULTIMODAL_ATTENTION_BACKEND kept pending #109 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
Add lobes/runtime/_detect.py: gathers host facts (nvidia-smi device name + compute_cap, /proc/meminfo total memory, /proc/device-tree/model, hostname) and resolves them via lobes.machines.detect() to a registered card name or UNKNOWN. Never reads nvidia-smi memory.used/memory.total (reports [N/A] on unified-memory boards); total memory comes from /proc/meminfo instead. Every probe is injectable and degrades to None on failure, never raising out of detect_card(). Stdlib only, no torch import. No registry extension was needed — name_markers already distinguish spark (GB10, sm_121) from thor (sm_110). 29 new tests in tests/test_detect.py; full suite 1260 passed / 14 skipped (baseline 1231 passed / 14 skipped).
…achines registry Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
Convert lobes/profiles.py into a package (lobes/profiles/) so the new
per-role fleet schema can live alongside the untouched legacy
WorkloadProfile/MachineProfile surface:
- lobes/profiles/schema.py: frozen Profile/RoleProfile dataclasses covering
cortex/senses/embedder/reranker, each with feasible/model plus the seven
machine knobs (gpu_mem_util, max_model_len, quantization, kv_cache_dtype,
attention_backend, enforce_eager, max_num_seqs). Unknown role/knob names
raise ModelGearError on load rather than being silently dropped.
- lobes/profiles/loader.py: resolves built-ins (packaged via
importlib.resources) + operator-defined profiles discovered under
<deploy-dir>/profiles/<name>.toml, with operator profiles overriding a
built-in of the same name. Nothing mutates a loaded Profile.
- lobes/profiles/builtin/{spark,thor}.toml: spark reproduces today's fleet
compose template values exactly (util 0.30/0.14/0.06/0.06, cortex 131072,
senses 32768, embed/rerank 8192); thor shares that baseline and the loader
overlays its 4 validated sm_110 divergences (cortex kv_cache_dtype=auto;
embedder+reranker attention_backend=TRITON_ATTN; reranker
enforce_eager=true) live from lobes.machines' thor CardStrategy/SM_110
trait, so the divergent values stay single-sourced instead of being
re-typed into the TOML.
Deviation: ships TOML, not the plan's spark.yaml/thor.yaml — the repo is
stdlib-only (dependencies = [], no YAML parser) and requires-python >=3.12
guarantees tomllib, so TOML is a zero-new-dependency stand-in for the same
data shape.
All 1231 pre-existing tests pass unmodified; 29 tests added (1260 passed, 14
skipped total). lobes/profiles/__init__.py re-exports the new schema/loader
API alongside the full legacy surface (MachineProfile, MACHINE_PROFILES,
detect_machine, resolve_serve_config, ...).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…operator override Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
detect card -> pick profile -> render .env: `lobes init` now resolves a fleet Profile (lobes.profiles, t1) via the injected host-card detector (lobes.runtime._detect, t2) and writes its per-role knobs into .env through a new pure lobes/profiles/render.py mapping (role -> PRIMARY_/ MULTIMODAL_/EMBED_/RERANK_ env prefix, knob -> env suffix, enforce_eager bool -> --enforce-eager/--no-enforce-eager flag token, feasible=False -> a <PREFIX>_FEASIBLE=false marker for a later task to honor). - --profile <name> overrides detection; forcing a profile onto a card it wasn't validated for (or an undetected one) warns on stderr but proceeds. - An UNKNOWN card with no --profile refuses (ModelGearError, EXIT_USER_ERROR) naming every detected fact and requiring --profile — never silently falls back to spark. Refuses before anything is written, dry run or apply. - Profile resolution/rendering is fleet-only; --single is unaffected (no detection call at all). - _apply_profile_env skips a key whose resolved value numerically matches what write_scaffold already copied from env.example, so a profile that agrees with the shipped template (e.g. spark on a fresh fleet .env, or thor's 3 unchanged knobs) doesn't reformat an untouched default. lobes/cli/_runtime_ops.py gains resolve_init_profile(), the shared detect+resolve+warn/refuse glue, with an injectable detect_fn (default resolved at call time so tests monkeypatch lobes.runtime._detect.detect_card directly, matching this repo's existing probe-neutralisation idiom). Verified on this Thor box: a bare `lobes init` (real detection, no flags) picks the thor profile.
…-> .env), --profile override, UNKNOWN refuses Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
Pin profile_env(resolve_profile(name)) for every shipped profile (spark,
thor) plus the fleet compose template's ${VAR:-default} surface as
committed goldens, diffed byte-for-byte. Rendering is pure (no host state,
no GPU, no subprocess — a purity meta-test proves it), so this is the CI
enforcement that a change scoped to one machine's profile/trait can't
silently alter another machine's rendering: touching thor.env's source
without also updating spark.env or template-defaults.env fails the suite.
tests/goldens/regen.py regenerates all three; tests/goldens/README.md
documents the "diff is the review surface" contract.
…diffed (cross-machine no-breakage guard) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
Extends issue #92's "advertised implies reachable" invariant to the HARDWARE dimension: a role a per-machine profile declares infeasible (RoleProfile.feasible=False) is never advertised ready in GET /capabilities / lobes capabilities, never listed in GET /v1/models, and a POST addressed to it 4xx's (role_infeasible) instead of being silently routed to a different gear. Channel: <PREFIX>_FEASIBLE=false env vars (PRIMARY_FEASIBLE / MULTIMODAL_FEASIBLE / EMBED_FEASIBLE / RERANK_FEASIBLE — see lobes.gateway._config.FEASIBLE_ENV), mirroring the existing <PREFIX>_MAX_MODEL_LEN served-context overlay convention. Computed once in build_config() onto RoutingTable.infeasible; every consumer (infeasible_owner, list_models_payload, lobes.roles.build_role_registry) reads it off the table already being threaded through — no parallel machinery, no new parameter on any existing public builder. RoleInfo gains a `feasible` field, clamping `ready` to False whenever a role's backend is hardware-infeasible, on top of the existing loaded/endpoint clamps — proven even when a live backend_ready signal reports the backend healthy. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…, gateway 404 role_infeasible) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
… + validation status Task t5 — lobes doctor/status now report: - Detected card (name or UNKNOWN) with device_name, compute_capability, total_memory_gb - Active profile from .env (persisted by init) - Whether profile is validated for the detected card - WARNING when card is UNKNOWN or profile doesn't match detected card Changes: 1. init: persist LOBES_PROFILE=<name> to .env (one-line hook) 2. doctor: add machine_profile section to output (text + --json) 3. status: add profile field to output (fleet + single, text + --json) 4. tests: comprehensive coverage with injected detection facts All 1388 tests pass (baseline 1383 + 5 new). doctor/status remain read-only (no filesystem writes in code path). Smoke test on Thor: correctly detects sm_110 compute capability, 122.8 GB memory. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…d/unvalidated Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…grade (t15) Vendors main's fleet/single docker-compose.yml + env.example (commit 7fa624d) into tests/fixtures/upgrade_compat/ as the "previous version" scaffold fixture, then drives status/serve/stop/init dry-run against a temp deployment dir built from it: every read/dry-run path succeeds, none writes (hashed before/after), and every ${VAR} main's template consumed is still honoured by the current template (the rename tripwire). init --apply without --force over the old scaffold still refuses rather than silently rewriting; --force --apply remains the only, explicit write path. No source changes were needed — t3/t4 stayed additive as claimed.
…rking, env names stable Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…refusing (t14)
resolve_init_profile no longer raises on an UNKNOWN card with no --profile.
It now resolves a new built-in 'base' profile (Qwen/Qwen3.5-4B cortex at
32K/util-0.30, senses infeasible, the standard 0.6B embed+rerank gears — no
27B anywhere) and returns a loud warning naming device_name/compute_capability/
total_memory_gb and the assumption made, so 'lobes init' proceeds (dry-run
prints the plan; --apply writes) instead of half-refusing. Recognised cards
(spark/thor, or an explicit --profile) are byte-identical — verified via
tests/goldens/{spark,thor,template-defaults}.env, unchanged by this commit.
Replaces the refuse branch t4 built; the broader "small default on every
card" change stays deferred to issue #107.
…e (no 27B) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…y gear's served name On a base-profile box the primary gear serves a 4B, but the template and env.example pinned the 27B id as the gateway's default model, routing no-model requests to a name nothing serves. Empty means _config.py's existing 'or primary.served_name' fallback wins, which preserves today's behavior on spark/thor byte-for-byte (their primary IS the 27B) and makes the default self-correct under any profile. Deliberate template change: template-defaults.env golden updated in the same commit (the t13 review surface); profile goldens untouched. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
… finding) On the real thinking-mode cortex the 16-token budget was consumed inside the <think> trace and content came back empty — the probe failed a correct model. Mirror lobes route's enable_thinking=false override (#93). Found running the probes against the live Thor fleet; with the fix all three probes pass on the tuned stack (rerank included — see #105/#106 note). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…d, verified against code
Document the full machine-profile surface:
- docs/machine-profiles.md (new): deep reference covering detection flow, profile
resolution, per-knob meanings and provenance (7 knobs × 4 roles), custom profiles,
goldens contract, support table, and Thor's four validated divergences
(kv_cache_dtype=auto, TRITON_ATTN pooling, enforce_eager reranker).
- lobes/explain/catalog.py: add _PROFILES entry with detection/resolution/knobs
overview; register ("profiles") and ("profile") shortcuts in ENTRIES.
- README.md: add support table (Spark load-tested, Thor load-tested with caveats,
base conservative fallback, Orin unvalidated) + link to machine-profiles.md.
- CLAUDE.md: add machine-profiles section with same support table, custom-profile
pointer, and references (detection, knobs, Thor divergences, custom profiles).
Validation: every knob value and env-var mapping verified against lobes/machines/
(thorton.py, _traits.py:SM_110), lobes/profiles/schema.py (ROLES, KNOB_NAMES),
lobes/profiles/render.py (role→prefix, knob→suffix), and lobes/profiles/builtin/
(TOML). All detection/machine/profile tests pass (14+18+29 total). afi cli doctor
. --strict: 26/26 checks pass. Claims traced to code (issue refs #105 #106 #107
#109 on Thor memory/hangs/eager/scale).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…d evidence The t11 draft carried claims the live validation contradicts; per the plan's added validation criterion, every corrected claim now traces to a real observation: - fp8-KV on Thor: the k_scale assert did NOT reproduce on the pinned nightly; fp8 boots with scale-1.0 warnings (accuracy risk, not crash). No probe ran under fp8 — said so instead of claiming one failed. - senses' TRITON_ATTN is the cross-machine Gemma head-size default, not a Thor divergence; the four real divergences are cortex kv auto, embed TRITON, rerank TRITON + eager. - Spark 'correctness probes pass' removed — the probes postdate the GB10 run and #106 is exactly that open question. - senses was NOT confirmed healthy in the clean-boot run — replaced the 'fully operational' caveat with the real one: the concurrent-boot memory race (4/4 failures), the drop_caches + primary-first workaround, and risk r7. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
PR Summary by Qodofeat: per-machine hardware profiles — detect the card, fit the box
AI Description
Diagram
High-Level Assessment
Files changed (63)
|
_render_text (doctor.py:266) tripped S3776 at complexity 19 — the loop over checks and the machine-profile section's chain of conditionals were both flattened into one function. Extract each into its own helper, _render_check_lines and _render_machine_profile_lines, and have _render_text just concatenate their output. Purely structural: same lines, same order, same JSON/exit-code contract. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…erving) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
_apply_machine_registry now builds the Profile with its constructor instead
of dataclasses.replace(), so the declared return type matches what static
analysis infers (S5886). test_duplicate_registration_rejected_without_replace
moves the machines.get("spark") setup outside the pytest.raises block so only
the register() call can raise (S5778).
…ises block Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…S3776) lobes/gateway/server.py:handle_post exceeded Sonar's cognitive-complexity ceiling (17 > 15) after t6's feasibility-gate additions. Extract four behavior-preserving helpers — _feasibility_response (shared 404 role_infeasible check), _resolve_tier (tier-alias branch: feasibility -> busy-shed -> served name), _resolve_plain_model (plain-id branch: unknown-id 404 -> feasibility -> served name), and _try_backends (the single-attempt per-backend loop) — so handle_post reads as a linear sequence of early returns instead of nested branching. Routing decisions, status codes, error bodies, and check ordering (feasibility outranks pressure; unknown-model checked before feasibility in the plain-id branch) are unchanged; the gateway test suite pins all of this and passes unmodified. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
… (behavior-preserving) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
- lobes/cli/_commands/assess.py (S3516, S3776): split cmd_assess's three branches (--probes / --preserve-thinking / default correctness) into helper functions, dropping cognitive complexity under the threshold. --probes now exits EXIT_ENV_ERROR when any probed role fails and EXIT_SUCCESS when all pass (mirrors lobes tunnel's status-driven exit code), giving cmd_assess genuinely different return values instead of always returning 0. --json/text payload shape is unchanged. - lobes/assess.py (S5713): drop the redundant json.JSONDecodeError from the embed-probe except tuple — it already derives from ValueError, which is caught in the same tuple. - tests/test_role_probes.py: pin the new exit-code contract (probe pass -> EXIT_SUCCESS, probe fail -> EXIT_ENV_ERROR in both --json and text mode). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
… redundant JSONDecodeError lobes assess --probes now exits EXIT_ENV_ERROR (2) when any probed role fails, 0 when all pass — same contract as tunnel's status exit. Payload shapes unchanged. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
Contributor
Author
|
/agentic_review |
Code Review by Qodo
Context used✅ Tickets:
🎫 verify reranker ORDERING on the DGX Spark GB10 — nothing currently checks it, it may be wrong there too 🎫 verify on the GB10 before t3 ships: is the fp8-KV crash checkpoint-driven, and is VLLM_ATTENTION_BACKEND truly dead on the pinned image? 🎫 Advertised per-role endpoint (:8000) 404s while ready=true — clients dialing it degrade instantly (lobes-cli#87 shape regressed)✅ Compliance rules (platform):
27 rules✅ Skills:
doc-test-alignment, cicd 1.
|
![qodo-code-review[bot]](https://avatars.githubusercontent.com/in/484649?s=60&v=4){kind=small}
…ofile names RoleProfile.from_dict validated key names only, so a TOML typo like feasible = "false" (a truthy string) silently passed construction and the renderer's truthiness check flipped a role feasible; now every knob's TYPE is checked (bool strictly for feasible, int/float-not-bool for numeric knobs, str-or-None for string knobs) and a bad value raises ModelGearError naming the role, knob, expected and got types. discover_operator_profiles() keyed profiles by the raw filename stem, so an operator file like Thor.toml never matched resolve_profile()'s .strip().lower() lookup and silently failed to override the builtin; keys are now normalised the same way, with a clear error on a post-normalisation collision (Thor.toml + thor.toml). Also fixed resolve_init_profile's --profile-vs-detected-card comparison to use normalised forms so a casing-only difference (--profile Spark on a detected "spark" card) no longer triggers a spurious mismatch warning. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…les (Qodo #2, #3) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
…AttributeError on malformed-but-200 probe bodies Finding 1 (Qodo): every --probes request travels through the gateway (RoleInfo.endpoint is always the gateway URL for the four gateway-fronted roles — lobes.roles._gateway_role sets endpoint = gateway). "cortex" is a resolvable gateway alias (lobes.catalog.TIER_ROLE -> tier_aliases), so the cortex probe now sends "cortex" as the model field instead of the concrete served id, exercising the same alias-resolution lane real callers use. embedder/reranker have no such gateway alias (TIER_ROLE only covers the generate lane) — sending those role names literally would 404 as an unknown model, so they keep sending the concrete served name. Finding 2 (Qodo): a 200-but-malformed body (e.g. "message" or a "data"/ "results" item is a bare string instead of a dict) made probe_cortex_correctness / probe_embed_correctness raise AttributeError from a `.get()` call, which their except tuples did not catch — aborting `lobes assess --probes` instead of reporting a structured FAIL. Added AttributeError to all three probes' except tuples (probe_rerank_correctness has no `.get()` call today but gets the same guard defensively, against a future refactor). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
… never raise on malformed 200s (Qodo #1, #4) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
Contributor
Author
|
Fixed in c9b72f8 (merged 65af1c3) — with a measured middle path. Evidence: every probe already routes through the GATEWAY (RoleInfo.endpoint is unconditionally the gateway URL in lobes/roles.py::_gateway_role), so for cortex the probe now sends the stable role alias (model=cortex), exercising the exact alias + hardware-infeasibility lane real callers use (#92/#81). For embedder/reranker the concrete served id is kept deliberately: the gateway alias table (tier_aliases() from catalog.TIER_ROLE) has no embedder/reranker entries, so a literal model=embedder would 404 as model_not_found. A _probe_model(role, info) helper encodes the rule and tests pin both sides. Extending pooling-role aliases to the gateway would be a separate feature — happy to file it if you think the alias contract should cover them too.
|
…ma.py S1192: the RoleProfile validator table repeated the "str or None" expected- type description four times; hoist it into a module-level _STR_OR_NONE constant and reuse it. S5799: the type-mismatch error message was built from two adjacent f-string literals on one line (implicit concatenation), which reads like an accidental split rather than an intentional one. It was cosmetic — a single string argument, not a tuple — so merge it into one explicit f-string; no behavior change.
…e f-string Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd
|
OriNachum
added a commit
that referenced
this pull request
Jul 14, 2026
…spec/brain-shapes Resolves the every-PR-bump collision on CHANGELOG.md / pyproject.toml / uv.lock by taking main's side (0.41.0); the branch's own 0.40.4 bump is dropped and re-applied on top as 0.41.1. #110 landed the #108 profile substrate this spec+plan composes on. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01U9ybFURhQxT9UpCgQNVL67
OriNachum
added a commit
that referenced
this pull request
Jul 14, 2026
#110 merged mid-flight (the #108 implementation, 0.41.0), so the frame and plan are re-cut against what actually shipped: the dropped-lobe contract is now flag-not-omit (capabilities feasible:false, /v1/models omits, generate lane 404 role_infeasible — c11/h3 rejected, c17/h13 recaptured and covered by t5); shapes own all six Colleague roles per user decision (stt/tts are first-class shape members over the audio overlay — c18, t1); shape data files are TOML matching lobes/profiles/builtin/; t1/t3/t5 instructions now name the landed modules instead of 'when it lands'. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01U9ybFURhQxT9UpCgQNVL67
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Implements the converged per-machine profiles plan (spec/plan: #108; 13 tasks, 4 waves, built via /assign-to-workforce with TDD-gated merges — suite grew 1188 → 1399, zero pre-existing tests broken).
What lands
lobes/machines/— per-chip strategy registry (CardStrategyper chip + sharedSM_110trait). LegacyMachineProfile/MACHINE_PROFILES/detect_machine()derived from it; adding a chip = one file + one registration line (proven by a synthetic-chip test).lobes/profiles/— schema (per-rolefeasible/model+ 7 knobs), TOML built-ins (spark,thor,base), loader with operator override, profile→env renderer. Thor's 4 divergences stay single-sourced in the registry and overlay at load time.lobes/runtime/_detect.py— card detection: device name + compute cap +/proc/meminfototal (never nvidia-smi memory fields —[N/A]on Thor). UNKNOWN is first-class.lobes initapplies the resolved profile;--profileforces with a warning; UNKNOWN warns and serves the conservative base profile (4B + pooling gears, no 27B — default install should assume a SMALL box, not a 128 GB Blackwell — conservative base profile + a weak default model #107's unknown-card slice).doctor/statusreport card + profile.<PREFIX>_FEASIBLE=false→ role omitted from capabilities, gateway 404role_infeasible(extends Advertised per-role endpoint (:8000) 404s while ready=true — clients dialing it degrade instantly (lobes-cli#87 shape regressed) #92).lobes assess --probes): known-answer / paraphrase-ranking / rerank-ordering; timeout = FAIL.tests/goldens/): byte-diffed per profile + template defaults, GPU-less — a Thor-side change cannot silently alter Spark's rendering (and vice versa).--force --applygated.docs/machine-profiles.md,lobes explain profiles, honest support tables.Live validation on the physical Thor (t10)
cudaErrorLaunchFailure) under TRITON_ATTN + eager. Senses unconfirmed (boots last; see below).k_scaleassert did not reproduce on the pinned nightly — uncalibrated fp8-KV is now an accuracy risk (scale-1.0 warnings), not a crash. verify on the GB10 before t3 ships: is the fp8-KV crash checkpoint-driven, and is VLLM_ATTENTION_BACKEND truly dead on the pinned image? #109 updated.drop_cachesafter teardown boots clean. Not expressible as a per-gear knob — recorded as plan risk r7, follow-up work.Deliberately NOT in this PR
MULTIMODAL_ATTENTION_BACKENDenv deletion — gated on the GB10 check (verify on the GB10 before t3 ships: is the fp8-KV crash checkpoint-driven, and is VLLM_ATTENTION_BACKEND truly dead on the pinned image? #109).🤖 Generated with Claude Code
https://claude.ai/code/session_01TJc5yvfweHP2AEccKNeaVd